Image courtesy of Dr. Zihao Wang
University of Tennessee at Chattanooga Assistant Professor Zihao Wang is leading a research collaboration that has achieved a significant breakthrough in interpretable 3D image modeling.
Wang, who joined the UTC Department of Computer Science and Engineering faculty in 2024, partnered with researchers from the French Institute for Research in Computer Science and Automation to develop a lightweight artificial intelligence model capable of learning to disentangle shape and appearance in images.
Their paper, “Multi-energy Quasi-Symplectic Langevin Inference for Latent Disentangled Learning,” was recently accepted by IEEE Transactions on Image Processing—a monthly peer-reviewed journal published by the Institute of Electrical and Electronics Engineers. The journal is considered one of the field’s most respected sources for research in image and signal processing.
Dr. Zihao Wang
Wang and his collaborators tackled a long-standing challenge in 3D image modeling: how to build models that are lightweight, interpretable and still deliver high-quality generative performance. Traditional deep learning approaches often achieve only two of those goals at once, he said.
The study introduced a new computational framework called the Langevin Variational Autoencoder (Langevin-VAE), which helps computers better understand the difference between an object’s shape and its surface details in medical images.
By using a quasi-symplectic integrator—a method that simplifies complex calculations—the model “avoids the expensive matrix calculations that typically slow down inference in high-dimensional data.”
“Our goal was to make deep generative models both interpretable and efficient,” Wang explained. “By integrating energy-based inference, we enable the model to learn how shape and appearance evolve independently without any supervision.”
The research team showed that their model could accurately analyze and rebuild 3D images of the inner ear and heart using a neural network with just 1.7 million parameters—much smaller than most similar models.
Despite its compact design, the Langevin-VAE outperformed larger state-of-the-art methods in both generative quality and disentanglement of latent features.
Beyond medical imaging, Wang said the framework opens new possibilities for interpretable AI systems in 3D modeling, robotics and scientific visualization.
This work was supported by the Ruth S. Holmberg Grant for Faculty Excellence, the UTC Department of Computer Science and Engineering, and the French National Research Agency.
“Dr. Wang’s research reflects the core values we champion at UTC CECS, curiosity that drives discovery, critical thinking that solves complex problems, and communication that bridges global collaboration,” said Dr. Kumar Yelamarthi, dean of the College of Engineering and Computer Science. “His work creates value in medical imaging and connects disciplines in ways that advance both science and society. This is the kind of innovation that empowers our students and faculty to lead with purpose.”
Wang was recently named a Ruth S. Holmberg Grant for Faculty Excellence recipient, which supports faculty members who demonstrate the potential to advance scholarship, engage students in research or develop innovative curriculum. Each recipient receives up to $5,000 to support their proposal.
His funded project, titled “Develop a Cross-Modal AI Agent for Medical Image Computing,” builds upon his ongoing research in lightweight and interpretable AI for medical applications.
Learn more
UTC College of Engineering and Computer Science
Computer Science and Engineering
Image courtesy of Dr. Zihao Wang
